LanceDB

LanceDB is where agent memories tend to live, and changing the embedding model there conventionally requires a destructive reset — throwing away exactly the accumulated memory that made the agent useful.

pip install "rebasis[lancedb]"

URI

lancedb:///absolute/path/to/db#table_name

Column names

LanceDB tables do not agree on what the columns are called. rebasis looks for the vector column among vector, embedding, embeddings, vec; the id among id, _id, doc_id, pk; and the text among text, content, document, page_content. That covers what the common tutorials and the LangChain integration produce.

When it guesses wrong, say so explicitly:

lancedb:///path/to/db#documents?vector_column=embeddings&id_column=chunk_id

If it cannot identify a column, the error lists both what it tried and what the table actually has.

What LanceDB supports

Capability	Supported
Read vectors	yes
Read text	when a text column exists
Upsert vectors	yes
Metadata filter	yes
Dimension locked	no
In-place update	yes

Unlike Chroma, LanceDB does not lock the dimension — so a full migration to a different-dimensional model is possible here, and probe will tell you whether it is worth doing.

Streaming

LanceDB is Arrow-backed and pages naturally, so reads stream without any special handling. rebasis never materialises a collection: peak memory is a function of the batch, not of the table.